123 - Robust NLP, with Robin Jia
NLP Highlights
00:00
Certifiably Robust Training: A New Approach to Training
So basically, you can just take a same kind of simple model that like a convolutional neural network, train it to do some analysis with this certifiably robust training. And we get something like 75% accuracy at test time on adversarially chosen perturbations. So if we would want to do something that understands, for example, perturbations at the phrase level or even at the sentence level, I think that's possible, but you would need a network that kind of respects that structure. For instance, probably what you would want is some system that first kind of generates sentence encodings for each sentence individually and then uses those to compute the final answer.
Transcript
Play full episode
